JFrog: 'Critical' SQLite CVEs Were LLM-Written Slop, Not Real Flaws

Security researchers at JFrog have concluded that a batch of advisory reports claiming critical SQLite vulnerabilities, published over the past few days, is largely AI-generated fiction. The analysis was published on the company's research blog on July 30.

What the advisories claimed

A newly created GitHub repository published a set of SQLite vulnerability reports as part of more than 50 CVEs that JFrog believes were produced by a large language model, with a single exception. The National Vulnerability Database flagged the reports as critical, and CISA's automated data processing program agreed. The SQLite entries described six use-after-free flaws in functions such as sqlite3ExprDelete(), jsonBlobEdit(), and jsonRemoveFunc(), scored between 7.5 High and 9.8 Critical.

Why the claims collapsed

Checking the reports against the official source code, the researchers found that the cited lines either did not exist in the affected versions or referred to unrelated logic. Some advisories named functions that do not exist; others listed fixed versions that were never released. The proof-of-concept SQL was run verbatim against official builds compiled in isolated Docker containers under AddressSanitizer, and none of it triggered a crash. Not one of the CVEs appears on SQLite's own advisory page, which the researchers call a gold standard for tracking real flaws. An AI-text detector flagged the combined advisories as machine-generated.

What the episode shows

Red Hat initially assigned CVE-2026-51302 a severity of 10.0 Critical, the maximum, before downgrading it to 7.6 High. The case illustrates how quickly unverified reports travel through the pipeline meant to protect users: they were scored and published before anyone reproduced a working exploit. Fake advisories waste maintainer time, distort vulnerability statistics, and erode trust in the CVE system at a time when AI-generated reports cost almost nothing to produce. JFrog's conclusion is simple: reproduction, not a severity score, is what separates a vulnerability from slop.